Spaces:
Running on Zero
Running on Zero
Fixed
Browse files
app.py
CHANGED
|
@@ -1,10 +1,54 @@
|
|
| 1 |
import os
|
| 2 |
import sys
|
|
|
|
|
|
|
|
|
|
| 3 |
import subprocess
|
| 4 |
import re
|
| 5 |
|
| 6 |
# =============================================================
|
| 7 |
-
# 1.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 8 |
# =============================================================
|
| 9 |
def bootstrap_llama_cpp():
|
| 10 |
print("Bootstrapping llama-cpp-python with CUDA for ZeroGPU...")
|
|
@@ -13,7 +57,6 @@ def bootstrap_llama_cpp():
|
|
| 13 |
api = HfApi()
|
| 14 |
files = api.list_repo_files("AIencoder/llama-cpp-wheels", repo_type="dataset")
|
| 15 |
|
| 16 |
-
# Priority 1: CUDA 12.4 / 12.6 wheels for ZeroGPU NVIDIA A10G
|
| 17 |
cuda_candidates = [
|
| 18 |
p for p in files
|
| 19 |
if p.endswith(".whl")
|
|
@@ -21,8 +64,6 @@ def bootstrap_llama_cpp():
|
|
| 21 |
and "basic" in p
|
| 22 |
and "/0.3.36/" in p
|
| 23 |
]
|
| 24 |
-
|
| 25 |
-
# Priority 2: CPU fallback
|
| 26 |
cpu_candidates = [
|
| 27 |
p for p in files
|
| 28 |
if p.endswith(".whl")
|
|
@@ -39,13 +80,17 @@ def bootstrap_llama_cpp():
|
|
| 39 |
subprocess.check_call([
|
| 40 |
sys.executable, "-m", "pip", "install", "--force-reinstall", "--no-deps", whl_path
|
| 41 |
])
|
| 42 |
-
print("llama-cpp-python
|
|
|
|
|
|
|
|
|
|
| 43 |
|
| 44 |
try:
|
| 45 |
from llama_cpp import Llama
|
| 46 |
-
print("llama-cpp-python
|
| 47 |
except (ImportError, RuntimeError, OSError):
|
| 48 |
bootstrap_llama_cpp()
|
|
|
|
| 49 |
from llama_cpp import Llama
|
| 50 |
|
| 51 |
import spaces
|
|
@@ -53,10 +98,8 @@ from typing import List
|
|
| 53 |
import gradio as gr
|
| 54 |
|
| 55 |
# =============================================================
|
| 56 |
-
#
|
| 57 |
# =============================================================
|
| 58 |
-
# In ZeroGPU, CUDA is ONLY active inside @spaces.GPU decorated functions.
|
| 59 |
-
# Loading Llama globally at startup fails because no GPU context exists yet.
|
| 60 |
MODEL_PATH = "Qwen3.5-0.8B.Q4_K_M.gguf"
|
| 61 |
llm = None
|
| 62 |
|
|
@@ -65,7 +108,6 @@ def get_llm():
|
|
| 65 |
if llm is None:
|
| 66 |
print(f"Initializing {MODEL_PATH} on ZeroGPU...")
|
| 67 |
try:
|
| 68 |
-
# Full GPU layer offload to ZeroGPU NVIDIA A10G
|
| 69 |
llm = Llama(
|
| 70 |
model_path=MODEL_PATH,
|
| 71 |
n_ctx=4096,
|
|
@@ -82,7 +124,7 @@ def get_llm():
|
|
| 82 |
n_gpu_layers=0,
|
| 83 |
verbose=False
|
| 84 |
)
|
| 85 |
-
print("Model loaded
|
| 86 |
return llm
|
| 87 |
|
| 88 |
SYSTEM_PROMPT = (
|
|
@@ -93,11 +135,10 @@ SYSTEM_PROMPT = (
|
|
| 93 |
)
|
| 94 |
|
| 95 |
# =============================================================
|
| 96 |
-
#
|
| 97 |
# =============================================================
|
| 98 |
@spaces.GPU(duration=30)
|
| 99 |
def gradio_chat(user_message, history):
|
| 100 |
-
# Model is loaded/invoked strictly INSIDE @spaces.GPU when CUDA context is granted
|
| 101 |
model = get_llm()
|
| 102 |
|
| 103 |
messages = [{"role": "system", "content": SYSTEM_PROMPT}]
|
|
@@ -121,7 +162,7 @@ def gradio_chat(user_message, history):
|
|
| 121 |
return response["choices"][0]["message"]["content"]
|
| 122 |
|
| 123 |
# =============================================================
|
| 124 |
-
#
|
| 125 |
# =============================================================
|
| 126 |
demo = gr.ChatInterface(
|
| 127 |
fn=gradio_chat,
|
|
|
|
| 1 |
import os
|
| 2 |
import sys
|
| 3 |
+
import glob
|
| 4 |
+
import ctypes
|
| 5 |
+
import site
|
| 6 |
import subprocess
|
| 7 |
import re
|
| 8 |
|
| 9 |
# =============================================================
|
| 10 |
+
# 1. Ensure CUDA Runtime Libraries (libcudart.so.12) are linked
|
| 11 |
+
# =============================================================
|
| 12 |
+
def fix_cuda_libs():
|
| 13 |
+
"""Expose bundled NVIDIA CUDA runtime libraries to dlopen for llama-cpp."""
|
| 14 |
+
candidate_dirs = []
|
| 15 |
+
for sp in site.getsitepackages():
|
| 16 |
+
nvidia_dir = os.path.join(sp, "nvidia")
|
| 17 |
+
if os.path.exists(nvidia_dir):
|
| 18 |
+
for root, dirs, _ in os.walk(nvidia_dir):
|
| 19 |
+
if "lib" in dirs:
|
| 20 |
+
candidate_dirs.append(os.path.join(root, "lib"))
|
| 21 |
+
for p in ["/usr/local/cuda/lib64", "/usr/local/cuda/lib"]:
|
| 22 |
+
if os.path.exists(p):
|
| 23 |
+
candidate_dirs.append(p)
|
| 24 |
+
|
| 25 |
+
if candidate_dirs:
|
| 26 |
+
current = os.environ.get("LD_LIBRARY_PATH", "")
|
| 27 |
+
os.environ["LD_LIBRARY_PATH"] = ":".join(candidate_dirs + ([current] if current else []))
|
| 28 |
+
|
| 29 |
+
# Preload libcudart and libcublas using RTLD_GLOBAL so libllama.so finds symbols immediately
|
| 30 |
+
for cdir in candidate_dirs:
|
| 31 |
+
for pattern in ["libcudart.so*", "libcublas.so*", "libcuda.so*"]:
|
| 32 |
+
for so_file in glob.glob(os.path.join(cdir, pattern)):
|
| 33 |
+
try:
|
| 34 |
+
ctypes.CDLL(so_file, mode=ctypes.RTLD_GLOBAL)
|
| 35 |
+
except Exception:
|
| 36 |
+
pass
|
| 37 |
+
|
| 38 |
+
def ensure_cuda_runtime():
|
| 39 |
+
fix_cuda_libs()
|
| 40 |
+
try:
|
| 41 |
+
ctypes.CDLL("libcudart.so.12", mode=ctypes.RTLD_GLOBAL)
|
| 42 |
+
print("libcudart.so.12 successfully loaded.")
|
| 43 |
+
except OSError:
|
| 44 |
+
print("libcudart.so.12 not found. Installing nvidia-cuda-runtime-cu12...")
|
| 45 |
+
subprocess.check_call([
|
| 46 |
+
sys.executable, "-m", "pip", "install", "nvidia-cuda-runtime-cu12", "nvidia-cublas-cu12"
|
| 47 |
+
])
|
| 48 |
+
fix_cuda_libs()
|
| 49 |
+
|
| 50 |
+
# =============================================================
|
| 51 |
+
# 2. Bootstrap prebuilt CUDA-enabled llama-cpp-python for ZeroGPU
|
| 52 |
# =============================================================
|
| 53 |
def bootstrap_llama_cpp():
|
| 54 |
print("Bootstrapping llama-cpp-python with CUDA for ZeroGPU...")
|
|
|
|
| 57 |
api = HfApi()
|
| 58 |
files = api.list_repo_files("AIencoder/llama-cpp-wheels", repo_type="dataset")
|
| 59 |
|
|
|
|
| 60 |
cuda_candidates = [
|
| 61 |
p for p in files
|
| 62 |
if p.endswith(".whl")
|
|
|
|
| 64 |
and "basic" in p
|
| 65 |
and "/0.3.36/" in p
|
| 66 |
]
|
|
|
|
|
|
|
| 67 |
cpu_candidates = [
|
| 68 |
p for p in files
|
| 69 |
if p.endswith(".whl")
|
|
|
|
| 80 |
subprocess.check_call([
|
| 81 |
sys.executable, "-m", "pip", "install", "--force-reinstall", "--no-deps", whl_path
|
| 82 |
])
|
| 83 |
+
print("llama-cpp-python package installed.")
|
| 84 |
+
|
| 85 |
+
# Make sure CUDA runtime is available in dynamic linker before importing llama_cpp
|
| 86 |
+
ensure_cuda_runtime()
|
| 87 |
|
| 88 |
try:
|
| 89 |
from llama_cpp import Llama
|
| 90 |
+
print("llama-cpp-python loaded successfully.")
|
| 91 |
except (ImportError, RuntimeError, OSError):
|
| 92 |
bootstrap_llama_cpp()
|
| 93 |
+
ensure_cuda_runtime()
|
| 94 |
from llama_cpp import Llama
|
| 95 |
|
| 96 |
import spaces
|
|
|
|
| 98 |
import gradio as gr
|
| 99 |
|
| 100 |
# =============================================================
|
| 101 |
+
# 3. ZeroGPU Lazy Model Loader
|
| 102 |
# =============================================================
|
|
|
|
|
|
|
| 103 |
MODEL_PATH = "Qwen3.5-0.8B.Q4_K_M.gguf"
|
| 104 |
llm = None
|
| 105 |
|
|
|
|
| 108 |
if llm is None:
|
| 109 |
print(f"Initializing {MODEL_PATH} on ZeroGPU...")
|
| 110 |
try:
|
|
|
|
| 111 |
llm = Llama(
|
| 112 |
model_path=MODEL_PATH,
|
| 113 |
n_ctx=4096,
|
|
|
|
| 124 |
n_gpu_layers=0,
|
| 125 |
verbose=False
|
| 126 |
)
|
| 127 |
+
print("Model loaded in CPU mode.")
|
| 128 |
return llm
|
| 129 |
|
| 130 |
SYSTEM_PROMPT = (
|
|
|
|
| 135 |
)
|
| 136 |
|
| 137 |
# =============================================================
|
| 138 |
+
# 4. Inference with ZeroGPU Allocation
|
| 139 |
# =============================================================
|
| 140 |
@spaces.GPU(duration=30)
|
| 141 |
def gradio_chat(user_message, history):
|
|
|
|
| 142 |
model = get_llm()
|
| 143 |
|
| 144 |
messages = [{"role": "system", "content": SYSTEM_PROMPT}]
|
|
|
|
| 162 |
return response["choices"][0]["message"]["content"]
|
| 163 |
|
| 164 |
# =============================================================
|
| 165 |
+
# 5. Gradio Interface (Exposing /chat API)
|
| 166 |
# =============================================================
|
| 167 |
demo = gr.ChatInterface(
|
| 168 |
fn=gradio_chat,
|